Is there currently an issue with VAPI? Our assistants are responding so slow and most of the time we have to make a follow up "hello" just to make it answer our previous question. Thanks!

Hi there! I understand you're experiencing slow responses with your VAPI Assistants and having to prompt with a follow-up "hello" to get responses. That's definitely frustrating! I don't have access to real-time status information about VAPI's systems right now, so I can't definitively say if there's a current outage. However, I can help explain potential causes and troubleshooting steps. There are several known issues that could be causing the slow responses: 1. "Deafness bug" - According to [Discord discussions](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1343612796717371393), VAPI has been experiencing a "deafness bug" where the assistant doesn't seem to hear user input. 2. Turn-taking behavior - The need for a follow-up "hello" might be related to how VAPI handles conversation turns. As mentioned in [another Discord thread](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1335054553250533439), "the user must provide input in order for it to be processed by the assistant for the subsequent response." 3. Configuration issues - Problems with Deepgram's endpointing value and startSpeakingPlan waitSeconds could be affecting response times. Setting "Deepgram's endpointing value to 300 and startSpeakingPlan waitSeconds to 0.8" has helped some users according to [this discussion](https://discord.com/channels/1211482211119796234/1211483291191083018/threads/1344177033625997322). To check if there's a current outage, I recommend visiting the VAPI status page. Several links to this page were found in the documentation, including [docs.vapi.ai](https://docs.vapi.ai/changelog/2025/2/25). For troubleshooting, you might want to: - Check your network infrastructure and latency - Review your VAPI configuration settings - Consider switching transcribers (from Deepgram to Whisper, Gladia, or Azure) - Check if you're exceeding concurrency limits

Could you share the call ID so I could take a look?

22f8fa47-c199-4744-9c2e-f155cbbc0dec 88920e79-25df-4beb-ae1f-acc9b52ffe6e d3907708-ea3b-4a45-a528-8098dda6844a

Hey SJ, What is happening is that when you say something or when you speak with the assistant, it goes to your knowledge base tool. When it goes to the knowledge base tool, it makes an inquiry. Then, you just Google "LLM model" to generate the response. However, at this time, as you are using your own API key, your API key does not have access to the model. Eventually, this results in an error, which causes the system to hang. Consequently, no response is generated, and nothing is voiced out. That's why you feel like the assistant is not responding, and you have to say "hello" again. I suggest you try another call by removing your own Google LLM API key. 🔵 19:22:07:420 Couldn't Complete Completion Request (#0, provider: google, model: gemini-1.5-flash, region: undefined, credential: false) Error: { "name": "ClientError", "stackTrace": {"code": 400, "status": "FAILED_PRECONDITION"}, "message": "\[VertexAI.ClientError\]: got status: 400 Bad Request. {\\"error\\":{\\"code\\":400,\\"message\\":\\"Project

3673231160

is not allowed to use Publisher Model `projects/angelic-bee-418405/locations/us-east5/publishers/google/models/gemini-1.5-flash`\\",\\"status\\":\\"FAILED_PRECONDITION\\"}}", "stack": "ClientError: \[VertexAI.ClientError\]: got status: 400 Bad Request. {\\"error\\":{\\"code\\":400,\\"message\\":\\"Project

3673231160

is not allowed to use Publisher Model `projects/angelic-bee-418405/locations/us-east5/publishers/google/models/gemini-1.5-flash`\\",\\"status\\":\\"FAILED_PRECONDITION\\"}}\\n at throwErrorIfNotOK (/app/node_modules/@google-cloud/vertexai/build/src/functions/post_fetch_processing.js:32:27)\\n at process.processTicksAndRejections (node:internal/process/task_queues:95:5)\\n at async generateContentStream (/app/node_modules/@google-cloud/vertexai/build/src/functions/generate_content.js:99:5)\\n at async GoogleRequest.request (/app/dist/libs/core/src/components/liveCallPipeline/streams/completions/requests/googleRequest.js:64:22)" }

Hi @Shubham Bajaj thanks for taking the time to look into it. However, we didn't set any Google LLM API Key on VAPI and when we created the tool.

@sj thanks for bringing this to my attention. The issue is most likely on our side. I'll check with the team and get back to you.